Speaker's meaning in the Gricean tradition is identified with the effect that, in performing a given utterance, the speaker intends, by means of the audience's recognition of this very intention, to produce in that audience (see Grice, Paul).
The GSV for a given utterance is the stacked mean supervector which is normalized as follows, widetilde{mathbf{m}}_{i} = sqrt{lambda_{i}} boldsymbol{Sigma}^{-1/2}_{i} mathbf{m}_{i} (19).
Even though misrecognized words may appear in any position of a given utterance, the partial fragments of the utterance passing the verification may have useful information for partial intention detection.
Motivated by the recent success of using DNNs in acoustic modeling for speech recognition, we adapt DNNs to the problem of identifying the language in a given utterance from its short-term acoustic features.
In the i-vector approach, assumed that we have obtained the i-vector w for a given utterance u (the expression of w is in Eq. (7), the process of calculating w is explained in [2]), the adapted mean vector of a Gaussian component can be written as: begin{array}rcl@ mathbf{M}_{c}=mathbf{m}_{c}+mathbf{T}_{c} mathbf{w} end{array} (5).
The Contextual component is restricted to representing only those aspects of the context of a given utterance which have a systematic influence on the form of that utterance.
Thus, the referent of 'the actual world' in a given utterance is simply the world of the speaker, just as the referent of an utterance of 'the present moment' is the moment of the utterance; likewise, an utterance of the form 'a is actual' indicates only that a shares the same world as the speaker.
Using these fundamental distributions and considering the graphical model represented by Figure 2, the likelihood of a given utterance with observations (o, g, c) can be factorized as (1).
For a given utterance, the speaker and channel variability dependent GMM supervector is denoted in Equation (1).
